Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-8x16c8-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_8x16c8__avx512vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 const int8_t* restrict a,
27 size_t a_stride,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 8);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 const int8_t* a0 = a;
46 float* c0 = c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
72 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 6) {
74 a5 = a4;
75 c5 = c4;
76 }
77 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
78 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
79 if XNN_UNPREDICTABLE(mr <= 6) {
80 a6 = a5;
81 c6 = c5;
82 }
83 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
84 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
85 if XNN_UNPREDICTABLE(mr != 8) {
86 a7 = a6;
87 c7 = c6;
88 }
89
90 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
91 XNN_FORCE_REALIZATION(vsign_mask);
92 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
93 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
94 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
95 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
96 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
97 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
98 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
99 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
100 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
101 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
102 // XNN_FORCE_REALIZATION(voutput_min);
103 // XNN_FORCE_REALIZATION(voutput_max);
104 do {
105 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
106 __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
107 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
108 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
109 __m512i vsum1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
110 __m512i vacc1x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 0));
111 __m512i vacc1x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 1));
112 __m512i vsum2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
113 __m512i vacc2x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 0));
114 __m512i vacc2x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 1));
115 __m512i vsum3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
116 __m512i vacc3x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 0));
117 __m512i vacc3x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 1));
118 __m512i vsum4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
119 __m512i vacc4x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 0));
120 __m512i vacc4x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 1));
121 __m512i vsum5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
122 __m512i vacc5x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 0));
123 __m512i vacc5x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 1));
124 __m512i vsum6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
125 __m512i vacc6x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 0));
126 __m512i vacc6x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 1));
127 __m512i vsum7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
128 __m512i vacc7x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum7x0123456789ABCDEF, 0));
129 __m512i vacc7x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum7x0123456789ABCDEF, 1));
130 w = (const int32_t*) w + 16;
131
132 size_t k = kc;
133 while (k >= 16 * sizeof(int8_t)) {
134 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
135 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
136 a0 += 16;
137 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
138 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
139 a1 += 16;
140 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
141 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
142 a2 += 16;
143 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
144 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
145 a3 += 16;
146 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
147 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
148 a4 += 16;
149 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
150 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
151 a5 += 16;
152 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
153 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
154 a6 += 16;
155 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
156 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
157 a7 += 16;
158
159 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
160 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
161 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
162 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
163 xnn_prefetch_to_l1((const int8_t*) w + 768);
164 xnn_prefetch_to_l1((const int8_t*) w + 832);
165
166 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
167 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
168 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
169 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
170 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
171 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
172 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
173 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
174 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
175 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
176 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
177 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
178 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
179 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
180 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
181 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
182 xnn_prefetch_to_l1((const int8_t*) w + 896);
183 xnn_prefetch_to_l1((const int8_t*) w + 960);
184 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
185 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
186 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
187 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
188 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
189 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
190 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
191 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
192 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
193 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
194 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
195 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
196 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
197 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
198 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
199 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
200
201 w = (const int8_t*) w + 256;
202 k -= 16 * sizeof(int8_t);
203 }
204
205 if (k != 0) {
206 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
207 a0 += 8;
208 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
209 a1 += 8;
210 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
211 a2 += 8;
212 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
213 a3 += 8;
214 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
215 a4 += 8;
216 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
217 a5 += 8;
218 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
219 a6 += 8;
220 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
221 a7 += 8;
222
223 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
224 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
225
226 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
227 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
228 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
229 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
230 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
231 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
232 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
233 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
234 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
235 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
236 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
237 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
238 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
239 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
240 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
241 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
242 xnn_prefetch_to_l1((const int8_t*) w + 896);
243 xnn_prefetch_to_l1((const int8_t*) w + 960);
244
245 w = (const int8_t*) w + 128;
246 k -= 8 * sizeof(int8_t);
247 }
248
249 // Add adjacent pairs
250 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
251 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
252 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
253 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
254 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
255 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
256 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
257 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
258 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
259 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
260 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
261 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
262 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
263 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
264 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
265 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
266 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
267 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
268 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
269 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
270 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
271 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
272 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
273 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
274 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
275
276 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
277 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
278 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
279 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
280 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
281 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
282 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
283 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
284
285 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
286 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
287 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
288 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
289 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
290 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
291 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
292 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
293
294 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
295 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
296 w = (const float*) w + 32;
297
298 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
299 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
300 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
301 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
302 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
303 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
304 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
305 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
306
307 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
308 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
309 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
310 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
311 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
312 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
313 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
314 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
315
316 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
317 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
318 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
319 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
320 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
321 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
322 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
323 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
324
325 if XNN_LIKELY(nc >= 16) {
326 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
327 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
328 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
329 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
330 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
331 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
332 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
333 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
334
335 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
336 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
337 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
338 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
339 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
340 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
341 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
342 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
343
344 c0 = (float*) ((uintptr_t) c0 + cn_stride);
345 c1 = (float*) ((uintptr_t) c1 + cn_stride);
346 c2 = (float*) ((uintptr_t) c2 + cn_stride);
347 c3 = (float*) ((uintptr_t) c3 + cn_stride);
348 c4 = (float*) ((uintptr_t) c4 + cn_stride);
349 c5 = (float*) ((uintptr_t) c5 + cn_stride);
350 c6 = (float*) ((uintptr_t) c6 + cn_stride);
351 c7 = (float*) ((uintptr_t) c7 + cn_stride);
352
353 nc -= 16;
354 } else {
355 // Prepare mask for valid 32-bit elements (depends on nc).
356 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
357 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
358 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
359 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
360 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
361 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
362 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
363 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
364 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
365 nc = 0;
366 }
367 } while (nc != 0);
368 }